Automatic Segmentation Method for Sphenoid Subregions in Cross-Scale Topologically Constrained Dual-Path Segmentation Networks

Abstract

An automatic segmentation method for sphenoid bone subregions using a cross-scale topologically constrained dual-path segmentation network is proposed. First, cranial CT or CBCT images are resampled, intensity normalized, and coarsely localized to obtain the target volume region of the sphenoid bone. Then, the target volume region is input into a dual-path encoding network. The global anatomical path is used to extract low-resolution, high-semantic features such as the overall morphology, spatial layout, and relative positions of subregions of the sphenoid bone, while the local boundary path is used to extract high-resolution detail features such as thin bone plates, fracturing boundaries, and protrusion junctions. Finally, a cross-scale coupling fusion module interactively gates and fuses the global anatomical features and local boundary features at multiple scales to obtain fused features that combine overall structural information and edge detail information. This invention significantly improves the automatic segmentation performance of complex subregions of the sphenoid bone.

Description

Technical Field

[0001] This invention belongs to the field of medical imaging and neurosurgical assistance under computer graphics, and relates to a segmentation method for the complex contours of the sphenoid bone subregion. Background Technology

[0002] The sphenoid bone, located in the center of the skull base, is an "anatomical hub" connecting the anterior, middle, and posterior cranial fossae, and also a "convergence center" for key structures such as nerves, blood vessels, and the sinuses at the skull base. Its structural complexity limits the limitations of traditional macroscopic anatomical divisions. The sphenoid bone consists of multiple parts, including the body, greater wing, lesser wing, and pterygoid processes (medial / lateral plates). Not only is its shape irregular, but it is also closely adjacent to surrounding structures. Medially, it directly connects to the pituitary fossa and cavernous sinus (containing the internal carotid artery, oculomotor nerve, and trochlear nerve), while laterally it participates in forming the superior orbital fissure, foramen rotundum, and foramen ovale (through which branches of the trigeminal nerve and ophthalmic vein pass). Inferiorly, it communicates with the nasopharynx and the sphenoid sinus (the deepest and most variable structure among the sinuses). This characteristic of "multi-regional fusion and multi-structure nesting" means that the traditional macroscopic anatomical division of "sphenoid bone = body + wing + processes" cannot meet the needs of refined diagnosis and treatment regarding "substructural boundaries, spatial relationships, and functional attributes."

[0003] The sphenoid region is a high-incidence area for skull base tumors, vascular lesions, trauma, and inflammation. The effectiveness of its diagnosis and treatment directly depends on the accurate identification and localization of "subanatomical units." The crudeness of traditional regional divisions has become a bottleneck restricting clinical development. Specific needs are reflected in the following areas: 1. Neurosurgery: Core needs for surgical planning and risk control. 2. Radiology: Precision requirements for imaging diagnosis and radiotherapy target areas. 3. Otolaryngology: Diagnostic and treatment needs for sinus surgery and diseases communicating with the skull base.

[0004] High-resolution imaging equipment (such as 320-slice spiral CT, 3.0T MRI, and PET-CT) can clearly display the fine structures of the sphenoid bone (such as the skull base foramen at the 0.5mm level and sphenoid sinus mucosal thickening). Combined with three-dimensional reconstruction technology (such as volume rendering VR and multiplanar reconstruction MPR), "three-dimensional visualization" of the sphenoid bone anatomy is achieved, providing objective imaging evidence for the "boundary definition" of subregions. For example, CT bone windows can accurately delineate the "connection between the lesser wing of the sphenoid bone and the optic canal," and MRI enhanced scans can distinguish the spatial relationship between the "internal carotid artery and cranial nerves within the cavernous sinus." Traditional subregion segmentation relies on manual delineation by radiologists, which is time-consuming and has significant individual differences (the overlap of the same subregion delineated by different doctors is only 50%-70%). In recent years, deep learning technologies (such as U-Net, 3D ResNet, and Transformer models) have developed rapidly in medical image segmentation. They can achieve "automatic and high-precision segmentation" of the sphenoid bone subregion through the process of "training on labeled datasets - model learning anatomical features - automatic segmentation of subregions". At the same time, they support multimodal image (CT+MRI) fusion segmentation, which solves the problem of incomplete display of "soft tissue-bone structure" by single modal images.

[0005] The research background of sphenoid bone subregion segmentation is essentially driven by the contradiction between the "need for clinical precision" and the "anatomical complexity and technical limitations": on the one hand, the diagnosis and treatment scenarios such as skull base tumors and radiotherapy urgently require the "quantification and standardization" of sphenoid bone substructures; on the other hand, breakthroughs in medical imaging and artificial intelligence technologies have provided feasible tools for subregion segmentation. The core objective of current research is to establish a standardized subregion system that integrates "anatomical morphology, clinical function, and individual variation," ultimately achieving a transformation from "empirical diagnosis and treatment" to "precision diagnosis and treatment." Summary of the Invention

[0006] To overcome the shortcomings of existing sphenoid bone subregion segmentation methods, such as omission of small subregions, adhesion between adjacent subregions, blurred boundaries, and unstable topological structure of segmentation results, this invention provides an automatic sphenoid bone subregion segmentation method based on a cross-scale topologically constrained dual-path segmentation network for the segmentation and identification of complex structures in sphenoid bone subregions. This method accurately identifies the location and boundaries of each subregion of the sphenoid bone, thereby reducing the risk of damage to important brain organs during surgery.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: An automatic sphenoid bone subregion segmentation method using a cross-scale topologically constrained dual-path segmentation network is proposed. First, the head CT or CBCT images are resampled, intensity normalized, and coarsely localized to obtain the target volume region of the sphenoid bone. Then, the target volume region is input into a dual-path encoding network, where the global anatomical path is used to extract low-resolution, high-semantic features such as the overall morphology, spatial layout, and relative positions of subregions of the sphenoid bone, while the local boundary path is used to extract high-resolution detail features such as thin bone plates, fracturing boundaries, and protrusion junctions. Finally, through a cross-scale coupling fusion module, the global anatomical features and local boundary features are interactively gated and fused at multiple scales to obtain fused features that combine overall structural information and edge detail information.

[0008] Furthermore, the method includes the following steps: 1) Image acquisition and standardization: Acquire head CT or CBCT images of the object to be segmented, and preprocess the images; 2) Sphenoid bone region localization and cropping: The sphenoid bone region is coarsely localized in the preprocessed 3D image of the skull to obtain candidate regions of the sphenoid bone; a bounding box is constructed based on the candidate regions of the sphenoid bone, and expanded and cropped according to a preset boundary margin to obtain the volume data of the target region ROI of the sphenoid bone. 3) Dual-path feature encoding: The ROI body data of the sphenoid bone target region is input into a dual-path encoding network to extract global anatomical features and local boundary features respectively; 4) Cross-scale coupling fusion: The global anatomical features and local boundary features are interactively coupled and fused at multiple scale levels to obtain fused features; 5) Anatomical prototype query decoding: Multiple learnable anatomical prototype vectors are constructed based on each subregion of the sphenoid bone, and the fusion features are queried and matched using the anatomical prototype vectors to generate spatial response maps of each subregion of the sphenoid bone; prototype competition normalization and mask reconstruction are performed on each spatial response map to obtain the initial segmentation results of each subregion of the sphenoid bone. 6) Topology constraint refinement: The initial segmentation results are input into the topology constraint refinement module, and the initial segmentation results are corrected for structural consistency based on the prior anatomical relationships between the subregions of the sphenoid bone. 7) Output the results: Output the final segmentation results and map the segmentation results back to the original image space to form a sphenoid bone subregion label map or a three-dimensional reconstruction result.

[0009] Compared with existing technologies, this invention significantly improves the automatic segmentation performance of complex subregions of the sphenoid bone by constructing a global anatomy and local boundary collaborative modeling mechanism, combined with subregion prototype-guided decoding and topological consistency refinement strategies. It is especially suitable for fine segmentation tasks of skull base bony structures that are small in size, have thin boundaries, and have complex adjacency relationships. Attached Figure Description

[0010] Figure 1 This is a flowchart of a segmented network. Detailed Implementation

[0011] The invention will now be further described with reference to the accompanying drawings.

[0012] Reference Figure 1 An automatic segmentation method for sphenoid bone subregions in a cross-scale topologically constrained dual-path segmentation network, the method comprising the following steps: 1) Image acquisition and standardization: Acquire head CT or CBCT images of the object to be segmented, and preprocess the images; the preprocessing includes: voxel resampling of the original three-dimensional images to ensure that different samples meet a uniform spatial resolution; truncation and normalization of image grayscale to ensure that the intensity distribution of bony structures falls within a preset range; and noise suppression processing of the images to reduce the impact of artifacts and high-frequency interference on subsequent segmentation. 2) Sphenoid bone region localization and cropping: The sphenoid bone region is coarsely localized in the preprocessed 3D image of the skull to obtain candidate regions of the sphenoid bone; a bounding box is constructed based on the candidate regions of the sphenoid bone, and expanded and cropped according to a preset boundary margin to obtain the volume data of the target region ROI of the sphenoid bone. 3) Dual-path feature encoding: The ROI body data of the sphenoid bone target region is input into a dual-path encoding network to extract global anatomical features and local boundary features respectively; 4) Cross-scale coupling fusion: The global anatomical features and local boundary features are interactively coupled and fused at multiple scale levels to obtain fused features; the fusion process is as follows: generating structural position guiding weights based on global anatomical features; generating boundary confidence weights based on local boundary features; performing cross-gating modulation and fusion on the global anatomical features and local boundary features to output multi-scale fused features that have both overall semantic information and local boundary information; 5) Anatomical prototype query decoding: Multiple learnable anatomical prototype vectors are constructed based on each subregion of the sphenoid bone, and the fusion features are queried and matched using the anatomical prototype vectors to generate spatial response maps of each subregion of the sphenoid bone; prototype competition normalization and mask reconstruction are performed on each spatial response map to obtain the initial segmentation results of each subregion of the sphenoid bone. 6) Topological Constraint Refinement: The initial segmentation results are input into the topological constraint refinement module. Based on the prior anatomical relationships between the subregions of the sphenoid bone, the initial segmentation results are corrected for structural consistency. The structural consistency correction process includes: adjacency consistency constraint; boundary exclusion constraint; connectivity constraint; thereby obtaining the final segmentation results of each subregion of the sphenoid bone. 7) Output the results: Output the final segmentation results and map the segmentation results back to the original image space to form a sphenoid bone subregion label map or a three-dimensional reconstruction result.

[0013] In this embodiment, learnable anatomical prototype vectors are established for each subregion of the sphenoid bone, and the fusion features are queried and matched using the anatomical prototype vectors during the decoding stage to generate spatial response maps corresponding to each subregion, thereby achieving accurate identification of the sphenoid bone body, greater wing, lesser wing, pterygoid process and other subregions; by introducing a prototype competition mechanism, the responses of multiple subregions at the same spatial location can be competitively normalized, effectively reducing misclassification and overlap between adjacent subregions.

[0014] Based on the initial segmentation results, a sphenoid bone subregion topological constraint refinement module is introduced. This module applies adjacency consistency constraints to the segmentation results based on the prior adjacency relationships between sphenoid bone subregions; applies boundary repulsion constraints based on the boundary relationships between adjacent subregions; and applies connectivity constraints based on the connectivity structure that each subregion should possess. This reduces unreasonable adhesions, breaks, and fragmentation, making the final output results more consistent with the actual anatomical structure. The detailed mechanisms of each module are as follows: The dual-path anatomical-boundary co-coding module comprises two parallel feature extraction branches. The global anatomical path employs a hierarchical encoding approach with a large receptive field to perform multi-scale semantic modeling of the input sphenoid bone region image, obtaining deep features representing the overall structural layout and subregional spatial relationships. The local boundary path uses a strategy of less downsampling or high-resolution preservation, and introduces edge-sensitive convolutions or gradient enhancement units to highlight the local responses of bony edges, thin structures, and subregional border areas. At each scale layer, gated coupling units are used to interactively fuse the features output from the two paths. Specifically, the global path generates structure-guided weights, and the local path generates boundary confidence weights. Both modulate the fused features, ensuring that the fused result contains both high-level anatomical semantics and preserves low-level spatial details.

[0015] Anatomical Prototype Query and Decoding Module: This module first constructs learnable prototype vectors for each subregion of the sphenoid bone. Each prototype vector represents the typical morphological semantics and spatial attributes of the corresponding subregion. During the decoding process, the fused spatial features are matched with the prototypes of each subregion to generate the response distribution of each prototype across the entire volume data, thus obtaining candidate segmentation regions for each subregion. Furthermore, to avoid overlapping activations of multiple subregions at the same location, a prototype competition mechanism is introduced. The responses of different prototypes at the same spatial location are normalized and competed for, so that the prototype of the subregion that best matches the anatomical attributes of that location receives a higher response, while prototypes that do not match are suppressed. Through this "prototype query—response generation—competitive allocation" approach, structured decoding for sphenoid bone subregions is completed.

[0016] The topology constraint refinement module constructs a subregion adjacency topology map based on the prior anatomical relationships of the sphenoid bone subregions and uses this map to correct the relationships in the initial segmentation results. Specifically, firstly, it detects abnormal contact in the predicted results based on the expected adjacency relationships of each subregion and penalizes contact between subregions that should not be directly adjacent; secondly, it applies boundary repulsion constraints to the boundary regions of adjacent subregions to suppress large-scale overlap between the two categories near the interface; and thirdly, it applies connectivity constraints to the connectivity structure of each subregion to reduce the probability of fragmented small regions or unreasonable breaks. In implementation, the initial segmentation map and intermediate semantic features can be input into the relationship inference unit to remodel the spatial relationships between subregions and output the corrected structurally consistent segmentation results.

[0017] Furthermore, this invention employs a multi-task joint training approach, simultaneously outputting sub-region segmentation maps, boundary maps, and topology maps. It also performs joint optimization using segmentation loss, boundary loss, prototype consistency loss, topology loss, and connectivity loss to improve the accuracy, stability, and structural consistency of sphenoid bone sub-region segmentation.

[0018] The loss function and evaluation metric for this embodiment are as follows: Dice coefficient ( Dice Coefficient F1 score, also known as the F1 score, is used to measure the degree of overlap between two sets. It ranges from 0 to 1, with a value of 1 indicating complete overlap and a value of 0 indicating no overlap. (1); Jaccard Index ( Jaccard Index Similar to the Dice coefficient, it also measures the degree of overlap between two sets, but is expressed as the ratio of the size of the intersection of the sets to the size of the union of the sets; (2); Precision rate ( Precision ): This measures the positive predictive value of a classifier, i.e., the proportion of true positives among all positive predictions; (3); Recall rate ( Recall ): Also known as sensitivity or true positive rate, it measures the proportion of all actual positive samples that are correctly predicted as positive; (4); Accuracy ( Accuracy ): refers to the overall correct classification rate of the classifier, that is, the proportion of correct predictions among all prediction results; (5); in, TP To predict the correct positive class portion, TN To predict the correct negative class part, FP For the more predicted positive class part, FN This refers to the missing positive class portion.

[0019] The embodiments described in this specification are merely examples of implementations of the inventive concept and are for illustrative purposes only. The scope of protection of this invention should not be considered limited to the specific forms described in these embodiments; rather, it extends to equivalent technical means conceived by those skilled in the art based on the inventive concept.

Claims

1. An automatic segmentation method for sphenoid bone subregions in a cross-scale topologically constrained dual-path segmentation network, characterized in that, First, the head CT or CBCT images are resampled, intensity normalized, and coarsely localized to obtain the target volume region of the sphenoid bone. Then, the target volume region is input into a dual-path encoding network. The global anatomical path is used to extract low-resolution, high-semantic features, including the overall morphology, spatial layout, and relative positions of subregions of the sphenoid bone. The local boundary path is used to extract high-resolution detailed features, including thin bone plates, pit and cleft boundaries, and protrusion junctions. Afterward, through a cross-scale coupling fusion module, the global anatomical features and local boundary features are interactively gated and fused at multiple scales to obtain fused features that contain both overall structural information and edge detail information.

2. The automatic segmentation method for sphenoid bone subregions in cross-scale topologically constrained dual-path segmentation networks as described in claim 1, characterized in that, The method includes the following steps: 1) Image acquisition and standardization: Acquire head CT or CBCT images of the object to be segmented, and preprocess the images; 2) Sphenoid bone region localization and cropping: The sphenoid bone region is coarsely localized in the preprocessed 3D image of the skull to obtain candidate regions of the sphenoid bone; a bounding box is constructed based on the candidate regions of the sphenoid bone, and expanded and cropped according to a preset boundary margin to obtain the volume data of the target region ROI of the sphenoid bone. 3) Dual-path feature encoding: Input the ROI body data of the sphenoid bone target region into the dual-path encoding network to extract global anatomical features and local boundary features respectively.

3. The automatic segmentation method for sphenoid bone subregions in cross-scale topologically constrained dual-path segmentation networks as described in claim 1, characterized in that, The method further includes the following steps: 4) Cross-scale coupling fusion: The global anatomical features and local boundary features are interactively coupled and fused at multiple scale levels to obtain fused features; 5) Anatomical prototype query decoding: Multiple learnable anatomical prototype vectors are constructed based on each subregion of the sphenoid bone, and the fusion features are queried and matched using the anatomical prototype vectors to generate spatial response maps of each subregion of the sphenoid bone; prototype competition normalization and mask reconstruction are performed on each spatial response map to obtain the initial segmentation results of each subregion of the sphenoid bone.

4. The automatic segmentation method for sphenoid bone subregions in cross-scale topologically constrained dual-path segmentation networks as described in claim 3, characterized in that, The method further includes the following steps: 6) Topology constraint refinement: The initial segmentation results are input into the topology constraint refinement module, and the initial segmentation results are corrected for structural consistency based on the prior anatomical relationships between the subregions of the sphenoid bone. 7) Output the results: Output the final segmentation results and map the segmentation results back to the original image space to form a sphenoid bone subregion label map or a three-dimensional reconstruction result.

5. The automatic segmentation method for sphenoid bone subregions in cross-scale topologically constrained dual-path segmentation networks as described in claim 2, characterized in that, In step 1), the preprocessing includes: voxel resampling of the original 3D image to ensure that different samples meet a uniform spatial resolution; truncation and normalization of the image grayscale to ensure that the intensity distribution of the bony structure falls within a preset range; and noise suppression of the image to reduce the impact of artifacts and high-frequency interference on subsequent segmentation.

6. The automatic segmentation method for sphenoid bone subregions in cross-scale topologically constrained dual-path segmentation networks as described in claim 3, characterized in that, In step 4), the fusion process is as follows: generating structural position guiding weights based on global anatomical features; generating boundary confidence weights based on local boundary features; performing cross-gating modulation and fusion on the global anatomical features and local boundary features to output multi-scale fusion features that combine overall semantic information and local boundary information.

7. The automatic segmentation method for sphenoid bone subregions in cross-scale topologically constrained dual-path segmentation networks as described in claim 4, characterized in that, In step 6), the structural consistency correction process is as follows: adjacency consistency constraint; boundary exclusion constraint; connectivity constraint; thereby obtaining the final segmentation results of each subregion of the sphenoid bone.

Images